Accurate and real-time prediction of umbilical component layout optimization based on convolutional neural network

Abstract

Umbilicals are an essential piece of control equipment in an underwater production system and are composed of cables, optical cables, steel pipes, and other components. Different component layouts result in umbilicals exhibiting different geometric characteristics and bearing capacities. Considering the geometric characteristics and mechanical properties of the umbilical components, a quantifiable optimization mathematical model for the layout of umbilical components is established in this work. The Differential Evolution-Generalized Lagrange Multiplier (DE-GLM) method is proposed to solve the mathematical model. We demonstrate, using mathematical examples, that the DE-GLM method can avoid the uncertainty of artificial experience layout design. Nevertheless, the increasing types, numbers, and layers of components lead to the time cost of DE-GLM calculation to be higher. To avoid the excessive consumption of time and artificial costs caused by traditional layout algorithms, a novel Efficient Pyramid-Scale Aggregation-U-Net (EPSA-U-Net) convolutional neural network is proposed to achieve real-time accurate prediction of umbilical component layout results. After discussing, in detail, the effects of hyperparameters such as optimization algorithms, learning rates, and loss functions, on the performance of the network, we demonstrate that the method not only automatically obtains an umbilical component layout design form that satisfies the performance requirements in real-time but also achieves an accuracy rate of 93.28% compared with other convolutional neural networks. As such, it could replace complex artificial operations, significantly improving the efficiency of umbilical component layout design within a negligible time frame. Simultaneously, it also provides an intelligent, reliable, and efficient reference for the layout design of other similar structures and equipment for marine engineering.